MODS: Fast and Robust Method for Two-View Matching

Dmytro Mishkin, Michal Perdoch, Jiri Matas, Karel Lenc

MODS: [pdf]
@article{Mishkin2015MODS, title = "MODS: Fast and robust method for two-view matching ", journal = "Computer Vision and Image Understanding ", volume = "", number = "", pages = " - ", year = "2015", note = "", issn = "1077-3142", doi = "http://dx.doi.org/10.1016/j.cviu.2015.08.005", url = "http://www.sciencedirect.com/science/article/pii/S1077314215001800", author = "Dmytro Mishkin and Jiri Matas and Michal Perdoch", keywords = "Wide baseline stereo", keywords = "Image matching", keywords = "Local feature detectors", keywords = "Local feature descriptors " }
[source][datasets] [paper]
MODS-WxBS: [pdf]
@InProceedings{Mishkin2015WXBS, author = {{Mishkin}, D. and {Matas}, J. and {Perdoch}, M. and {Lenc}, K. }, Booktitle = {Proceedings of the British Machine Vision Conference}, Publisher = {BMVA}, title = "{WxBS: Wide Baseline Stereo Generalizations}", year = 2015, month = sep,}
[poster] [source][datasets]

Description

MODS (Matching On Demand with view Synthesis) is algorithm for wide-baseline matching. It matches image with extreme view point changes, which are not matchable with previous state-of-art ASIFT matcher. MODS is fast for simple, robust on hard matching problems because of the use of progressively more time-consuming feature detectors and by on-demand generation of synthesized images that is performed until a reliable estimate of geometry is obtained.

Challenging image pairs sucessfully matched by MODS(-WXBS) where other methods fail.


Comparison of MODS to state-of-art matchers. Image pairs successfully matched reported
Matcher/Dataset VGG-Aff EF SymBench MMS GDB-ICP UW EVD EZD Lost in Past WxBS
#Image pairs 40 33 46 100 22 30 15 22 172 31
MODS 40 33 42 27 18 9 15 12 94 6
MODS-WxBS 40 33 43 82 22 9 15 n/a 107 12
ASIFT 40 23 27 18 15 6 5 8 62 1
DualBootstrap 36 31 38 79 16 n/a 0 4 16 0

WXBS-MODS is extension designed to handle also extreme appearance changes like visible-IR, day-night, summer-winter matching in addition to the wide geometric baseline.

As one application of MODS-WXBS, we have built place recognition engine for VPRiCE Place Recognition challenge, which showns best Precision and F1 score so far and outperformed CNN-based competitors.

Papers

[MODS] MODS: Fast and Robust Method for Two-View Matching.. D. Mishkin and M. Perdoch and J.Matas. CVIU, 2015 [pdf],
@article{Mishkin2015MODS, title = "MODS: Fast and robust method for two-view matching ", journal = "Computer Vision and Image Understanding ", volume = "", number = "", pages = " - ", year = "2015", note = "", issn = "1077-3142", doi = "http://dx.doi.org/10.1016/j.cviu.2015.08.005", url = "http://www.sciencedirect.com/science/article/pii/S1077314215001800", author = "Dmytro Mishkin and Jiri Matas and Michal Perdoch", keywords = "Wide baseline stereo", keywords = "Image matching", keywords = "Local feature detectors", keywords = "Local feature descriptors " }
[MODS-WxBS] WxBS: Wide Baseline Stereo Generalizations. D. Mishkin and M. Perdoch and J.Matas and K. Lenc. In Proc BMVC, 2015 [pdf],
@InProceedings{Mishkin2015WXBS, author = {{Mishkin}, D. and {Matas}, J. and {Perdoch}, M. and {Lenc}, K. }, Booktitle = {Proceedings of the British Machine Vision Conference}, Publisher = {BMVA}, title = "{WxBS: Wide Baseline Stereo Generalizations}", year = 2015, month = sep,}
[MODS-VPRiCE] Place Recognition with WxBS Retrieval. D. Mishkin and M. Perdoch and J.Matas, CVPR 2015 Workshop on Visual Place Recognition in Changing Environments [pdf],
@InProceedings{Mishkin2015VPRICE, author = {{Mishkin}, D. and {Perdoch}, M. and {Matas}, J. }, title = "{Place Recognition with WxBS Retrieval}", book = {CVPR 2015 Workshop on Visual Place Recognition in Changing Environments}, year = 2015, month = jun, }

Source code

Linux C++ source code, Windows C++ source code, Linux 64-bit binaries(old version). Sources code for MODS and MODS-WXBS are implemented in single package, which is congifurable via ini-files.
MODS-IVCNZ binaries (executable files, 64-bit Linux, Windows). 2013 year version, provided for results reproducability.

Datasets

The Extreme View Dataset. EVD is a set of 15 image pairs with ground truth homographies (image pairs - "adam", "graf" and "there" are from [1] and [2]).
The Extreme Zoom Dataset. EZD is a 6 image sets with incleasing zoom factor from general scene view to focusing on single detail.
The Wide (multiple) Baseline Dataset. 31 image pairs, simultaneously combining severel nuisance factors: geometry, illumination, modality, etc.
The Wide Single Baseline Dataset. 40 image pairs, grouped by nuisance factor: geometry, illumination, appearance, modality, with pre-detected local features by Hessian-Affine, Edge FOCI and MSER.
The map2photo dataset - 6 pairs, where one image is satellite photo and second - map of the same area.

References

[1] K. Cordes and B.Rosenhahn and J. Ostermann. Increasing the Accuracy of Feature Evaluation Benchmarks Using Differential Evolution. In Proc of SSCI, 2011.
[2] G. Yu, and J.-M. Morel ASIFT: An Algorithm for Fully Affine Invariant Comparison. IPOL vol. 2011. http://dx.doi.org/10.5201/ipol.2011.my-asift
[3] D. Mishkin and M. Perdoch and J.Matas. Two-View Matching with View Synthesis Revisited. In Proc. of IVCNZ, 2013, 436-441. [pdf]


Dmytro Mishkin,